Applying deep learning neural networks for automatic validation of forest management materials based on satellite imagery

The article discusses the use of deep learning neural networks to verify forest taxation indicators obtained during forest management. The authors have designed a technology for verifying forest management data based on the high-resolution optical satellite images processed applying neural networks. The initial data were images of the Melenkovskoye forestry district in the Vladimir region taken from Sentinel-2 satellites; atmospheric correction was performed with the Sen2Cor algorithm. The process of forest taxation interpretation of images was automated with the «Geotron. Forest Validation» software, a specialized geoinformation resource developed by the NC OMZ, FBU VNIILM and FGBU “Roslesinforg”. The TensorFlow and Keras libraries were used to train the neural network models; the training sample was based on the average pixel values within the boundaries of each plot for every channel. After training, the neural network determined forest characteristics for each section, and also gave the difference between the forest management documentation data and those calculated by the neural network; this allowed identifying deviations in the stock volume, forest density, and stands age. The model's operation demonstrated that a significant portion of inaccurate forest management data is typical for complex stands with several species in the composition, as well as for sparse maturing and mature stands with different species composition of the first tier. The study result showed that the accuracy of neural network training and operating directly depends on the of data amount, in particular, the number of sections for which training occurs. Acceptable results are obtained with a sections number exceeding 5,000 pieces. Considering this, the developed technology allows obtaining good results in validating forest management data on an area exceeding 10 thousand hectares, which is comparable to the average area of one forestry district in the central part of Russia.

1. Abadi M. TensorFlow: learning functions at scale. Proceedings of the 21st ACM SIGPLAN International Conference on Functional Programming: ICFP 2016. New York, 2016, p. 1.

2. Bogodukhov M.A., Bartalev S.A., Zharko V.O. Study of the possibilities of recognizing predominant forest species using regression assessment of the share of species by stock based on Sentinel-2 data and materials from sample plots. Current problems in remote sensing of the Earth from space: mat. of 22nd int. conf. Moscow, 2024, p. 159. (In Russ.)

3. Brandt M., Chave J., Li S., Fensholt R., Ciais P., Wigneron J.-P., Gieseke F., Saatchi S., Tucker C.J., Igel C. High-resolution sensors and deep learning models for tree resource monitoring. Nature Reviews Electrical Engineering, 2025, vol. 2, no. 1, pp. 13-26.

4. Ivanov S.V., Sidorenkov V.M., Achikolova Yu.S., Astapov D.O., Tishchenko P.A., Buryak L.V., Rybkin A.S. Possibilities of using data from the BRICS remote sensing satellite constellation to solve thematic issues of obtaining information on forest ecosystems. Forestry information, 2024, no. 4, pp. 42-55. (In Russ.)

5. Main-Knorn M., Pflug B., Louis J., Debaecker V., Müller-Wilm U., Gascon F. Sen2Cor for Sentinel-2. Proceedings of SPIE, 2017, vol. 10427, pp. 37-48.

6. Markov N.G., Machuka K.R. Deep learning models and methods for solving problems of forest resources remote monitoring. Bulletin of Tomsk Polytechnic University. Georesources engineering, 2024, vol. 335, no. 6, pp. 55-74. (In Russ.)

7. Marques T., Carreira S., Miragaia R., Ramos J., Pereira A. Applying deep learning to real-time UAV-based forest monitoring: Leveraging multi-sensor imagery for improved results. Expert Systems with Applications, 2024, vol. 245: Applying deep learning to real-time UAV-based forest monitoring, pp. 107-123.

8. Melnikov A.V., Polischuk Yu.M., Rusanov M.A., Abbazov V.R., Kochergin G.A., Kupriyanov M.A., Baysalyamova O.A., Sokolov O.I. Comparative analysis of neural network models for mapping forest cuttings based on summer space images. Scientific and Technical Bulletin of Information Technologies, Mechanics and Optics, 2024, vol.24, no. 5, pp. 806-814.

9. Sen2Cor. Science Toolbox Exploitation Platform // European Space Agency. URL: https:.step.esa.int/main/snap-supported-plugins/sen2cor/ (accessed June 10, 2025)

10. Sidorenkov V.M., Martynyuk A.A., Serezhkin A.V. Using Sentinel-2 satellite imagery data to assess the characteristics of forest vegetation in taiga forests. Ecological and biological foundations for increasing the productivity and sustainability of natural and artificially restored ecosystems: proceedings of the int. sci. conf. Voronezh, 2018, pp. 103-115. (In Russ.)

11. Sidorenkov V.M., Astapov D.O., Perfilyeva O.V., Ryabtsev O.V., Rybkin A.S. Determination of forest characteristics of pure pine stands based on Kanopus-V imagery data. Rocket and Space Instrumentation and Information Systems, 2022a, vol. 9, no. 2, pp. 36-43. (In Russ.)

12. Sidorenkov V.M., Astapov D.O., Rybkin E.S., Achikolova I.S., Ryabtsev O.V. Possibilities of using satellite imagery from the Meteor-M spacecraft to determine the quantitative and qualitative forests characteristics. Forestry information, 2022b, no. 2, pp. 5-12.

13. Zhang Z., He G., Wang X. A practical DOS model-based atmospheric correction algorithm. International Journal of Remote Sensing, 2010, vol. 31, pp. 2837-2852.